Special Project

"They didn't cut people, but the boring part of their work." AI has already entered the Belarusian business

Source: Olga Prokofyeva, photos by Vlad Borisovich, Alexander Ruzhechka, Anastasia Chubenko

It seems that business has already passed the stage where neural networks were shown to colleagues as an entertaining experiment: "Look, AI beat Excel." Now, neural networks are being entrusted with real work - analyzing conversations with customers, sorting documents, preparing presentations, and searching for errors in processes. But the more business delegates tasks to machines, the more often another question arises: do we need to stop them at some point? In a joint project with hoster.by, we discussed trends and also visited the offices of Belarusian companies to see where AI saves time and money, and where it is still just trying its strength.

"The emergence of AI caused a well-deserved wow effect," says Mikhail Prikhodko, head of the cloud solutions department at hoster.by. "There was an understandable temptation to think that since an accessible technology can solve one-time tasks almost magically, it will easily cope with business processes as well. Everything turned out to be a bit different. And now, AI optimism is being replaced by AI realism."

According to the expert, during the period of AI optimism, only a small part of companies perceived the technology not as a goal, but as a means. They identified tasks, gradually and pointwise began to integrate them into processes, controlled, and measured effectiveness.

"Many more companies perceived the topic as a 'hype', and in the background, they still went to 'play with ChatGPT', or joined the fake-it-till-you-make-it approach: soon we won't need designers, programmers, marketers - AI will do everything. Both of these categories were disappointed."

Now, the period of AI realism is coming, where extremes are being replaced by balanced decisions about the implementation of tools - what, why, and how we want to achieve. And the main thing - what benefits this will bring to the company.

Implementation of AI. Experience of "Alfa Bank"

Experience of "Aigenis"

Experience of SIPSIM

Experience of "Modum"

"The bank is becoming an AI agent for humans"

At "Alfa Bank", artificial intelligence was developed long before the current boom in generative models. Initially, these were machine learning technologies: models helped solve financial problems and assess risks, and their calculations were used in work processes. But with the spread of ChatGPT, the bank began to look for another application of AI - not just to calculate and analyze, but to become a helper for humans.

The first direction was to assist clients in chats. Simply connecting ChatGPT or another external service was not possible: in the banking sector, security and control over data are critical.

If a bank loses security, its products and services are no longer important, because they rely on the security of customers' money and data, which the bank guarantees, says Dmitry Nagorny, head of the data and artificial intelligence department at Alfa Bank.

Therefore, the bank decided to create its own solutions based on open models within a secure perimeter. Initially, they tried to fine-tune models on internal data, but the quality improved only slightly: the volume of corporate information turned out to be insufficient compared to the data of the entire internet, on which the models were initially trained. One of the solutions was the RAG technology, which allows the model to work with pre-prepared information, making the answers more accurate.

Thus, in 2024, a chatbot for clients appeared. It is positioned as an intelligent assistant: it consults on various topics, and if a person wants to talk to an operator, they can be switched to a specialist at any moment. The bot's work is evaluated according to the same criteria as the work of a support specialist, including customer satisfaction.

For us, the client experience is primary. Technology should be ready for this, says Dmitry.

In parallel, the bank began to develop a tool for employees - Alfa Assistant. This is a platform for working with generative artificial intelligence: you can not only communicate with the model in chat mode but also upload files, images, and create your own personal assistants based on your data. Currently, more than half of the bank's employees use Alfa Assistant.

The next step was to create specialized assistants for individual areas. For example, the HR assistant answers questions about internal procedures, and in the future, it should not only explain how to fill out a sick leave form but also do it independently (i.e., become a full-fledged AI agent).

For example, an employee says: I want to take a vacation from this date to that date. The assistant checks all the necessary data in accordance with the legislation and internal regulations, and that's it - the vacation application is ready. This is our goal - simplicity. To achieve it, we need to go a long way of research, data preparation, and development.

The legal assistant works with almost 13,000 internal bank documents, including accumulated practices, as well as taking into account external legislation. Its task is to help the lawyer find information faster and prepare answers.

• Blindly trusting the answer is not possible. An error is always possible, such is the peculiarity of the technology - the question is about the probability of this error. To reduce it, we work with the quality and updating of data, introduce control points, use model arbitration and other approaches. In legal consulting, an error can be very painful, so despite the technology, human participation is necessary, says Dmitry.

If the AI assistant answers questions and helps work with information, then the agent must be able to act independently. The first agents already exist, but now they are being combined into chains that can perform entire processes. For example, in lending, AI can help recognize documents, prepare materials for credit committees, and form contracts. Nevertheless, final control and decision-making remain with a person.

AI is gradually changing not only individual processes within the bank but also the way humans interact with financial services.

• An important feature of generative AI is its human-like nature: it changes the interface for interacting with bank services, when a person can write or say what they want and be understood, and receive a competent answer.

This approach also changes the understanding of where interaction with the bank begins. Let's imagine that a person needs to quickly transfer money. They say to Siri: "Transfer money to this person." The voice assistant understands who is being referred to and can find the phone number, contact the bank's financial agent, who performs the transfer by phone number and returns the result of the financial operation. As a result, the person contacted Siri and received an answer in real-time, without even suspecting the entire chain of processes.

• The technologies for such a scenario are largely ready. We are actively exploring the possibility of ensuring data security in such inter-agent interaction, which is key to making a decision to go down this path.

Our goal is to become an AI agent for humans and provide banking services, including in this form. To achieve this, we are transforming in such a way that the design of processes, products, and services is based on the capabilities of AI. We strive to build an ecosystem of agents around key processes and gradually interact not only with clients and employees but also with external agents.

However, in all these global changes, it is essential to remember that everything must be done in the interests of humans. The key to any company is its intellectual potential and satisfied clients and employees. No matter what new technology emerges, everything ultimately depends on how competently a person applies it.

Few have implemented AI, but it is present in almost every company

There is a paradox in that only a small part of companies implement AI-based solutions in their processes. However, in reality, AI is present in one form or another in almost every company, says Vladimir Pisan, system architect at hoster.by.

Enormous amounts of work information, and sometimes personal data, commercial secrets, and other restricted information, are being sent to conditional ChatGPT or Claude.

Few people think about the consequences when so much convenience can be obtained "here and now." However, the information sent to third-party services is stored on servers that belong to these services, not to the users. This data is used, among other things, for training models. If you ask a chat to create a fragment of a program, where does it get this code from, considering that it cannot create anything itself?

More and more managers are consciously seeking access to LLMs that can be worked with in their own isolated environment. This means a model where you can upload your own data, which will be processed in an accredited data center contour, and, moreover, in full compliance with the legislative requirements for personal data protection. And yes, you can still hear people say, "Is that really possible?"

"We only trust money to humans"

More than half of the employees at the investment company "Aigenis" already work with artificial intelligence in one way or another. The company's client base is growing, and so is the workload. Therefore, a year and a half ago, the company began looking for a solution using AI, says Elena Ersh, deputy director of "Aigenis".

Initially, they decided to try using neural networks in client work, sales, and marketing, but they needed to resolve the security issue.

For us, as for banks, it was crucial to comply with legislation and internal regulations when working with commercial information and personal data, and to ensure that sensitive information did not leave the protected perimeter of the company. The companies that offered us their solutions were unable to provide an option that fully satisfied us in terms of security. Therefore, the project was put on hold.

About six months later, Aigenis found a contractor and started with five AI agents, which took on routine tasks. Three of them work with the sales department: they evaluate managers' calls, help process objections, and automatically fill out a card in the CRM system after a conversation. The other two are used throughout the company - "Presentation Master" and "Meeting Secretary".

After the first positive experience, they organized their own hackathon: in eight weeks, they wanted to create 20 AI agents. As a result, 11 of them reached the stage of ready-made solutions, and they had to abandon three agents - they honestly admitted to themselves that AI was not needed in those cases.

The constant "bottleneck" in the company remained the issue of the time spent writing protocols, recording agreements, and describing tasks for subsequent control. Each department head had previously worked out a protocol form for themselves, determining what they wanted to get as output.

A similar story happened with presentations: the AI agent "Presentation Master" was loaded with a brand book, past materials, and design requirements.

Previously, we either created presentations ourselves or turned to an external designer. Now the agent creates the presentation completely independently, and our criterion for effectiveness is 30 minutes to the final version. Previously, this took hours.

From internal tasks, AI gradually moved on to working with clients. If you write questions in the company's Telegram channel, an AI agent will answer some of them. For this, a knowledge base was previously collected for it, and possible answer options were tested. If the question is complex or there is no answer in the knowledge base, the request is forwarded to a manager.

At the initial stages of development, we saw how AI would hallucinate if it didn't know the answer. Therefore, in the absence of a precise answer, the question should be sent to a manager. What is not in the database, the AI should not invent.

Another agent helps work with the client base. Client data is anonymized before being sent to the platform. Clients are divided into groups, for example, based on how long they have been purchasing securities, how much they have purchased, and the structure of their investment portfolio. The company analyzes the client's behavior under the influence of external events: changes in rates, exchange rates, marketing activities, and other factors.

Thus, we hope to track behavioral trends and form personalized offers for each client, taking into account how they reacted to different events in the past. For now, we are learning to analyze large arrays of disparate data, testing different hypotheses to identify recurring trends.

After several successful experiments, we decided to automate the morning news digest. And that's where it didn't work out.

Part of the information we couldn't parse for free. In some cases, the information was duplicated on several news channels, and some sources were simply inaccessible. We didn't get a complete picture, so we stopped developing the agent and continue to monitor the news manually.

A similar story happened with the agent that was supposed to collect information on deposit rates from different banks. The banks' data was not always updated in a convenient format for automatic collection.

Sometimes it's faster and better to do it manually than to create an agent, spend resources on it, and then get a result that doesn't meet our expectations, says Elena Ersh. For each agent, there are criteria for effectiveness - for example, the speed and accuracy of the response or the speed of preparing a report. If the result doesn't meet the required level, the project is put on pause or abandoned.

We also tried using AI for financial analysis, but it didn't impress us.

We have our own specifics, rules, and methodology for analyzing data. Template models don't work for us. However, we were able to extract information into dashboards. That is, the finished analysis can be quickly structured and displayed in the format of graphs and diagrams by AI.

But there is work that the company principially does not entrust to AI: deciding where to invest money.

Artificial intelligence can suggest a solution, but the choice is always left to a person - to accept it or not. When it comes to money, any financial decision involves taking risks and responsibility. I believe it's unacceptable to shift this responsibility to AI when we're dealing with large sums of money.

The neural network learns from past data, but the market can drastically change the rules of the game, says the interviewee.

As soon as a "black swan" event occurs or there is a change in exchange rates, rates, or prices, the market readjusts, and what worked before becomes ineffective and less profitable. AI can help calculate different scenarios, but making the decision about which scenario to follow is still a human prerogative.

AI is increasingly solving specific business problems

One of the most common scenarios for using AI in Belarus is a corporate closed chat, where employees can upload